Research On Small-scale Infrared Target Detection Algorithm Based On Deep Learning

The task of target detection is one of the most important tasks in computer vision,and it is also a classic research topic.The infrared imaging system relies on the infrared radiation and background radiation emitted by the object.Compared with the traditional detection system,it can penetrate smoke,fog,and other unfavorable atmospheric conditions during operation and can provide clear images without illumination.Because of these advantages,infrared detection systems have broad application prospects in military and civilian applications,such as early warning systems,maritime surveillance systems,automotive night vision systems,and so on.Therefore,the task of target detection for infrared images has important research significance and application value.However,there are many challenges in infrared image and infrared target detection tasks.The current research mainly faces the following difficulties:(1)The problem of minimum intrinsic features: Because infrared detection systems often have a large field of view and a long imaging distance,some infrared targets only take up a few picture elements in the image and lack texture or shape feature.The lack of texture or shape features often leads to unreliable target detection,especially in the current deep learning network,small targets are easily submerged by complex backgrounds.(2)The contradiction between semantic information and resolution: In the convolution process of deep networks,the original image will gradually become smaller with the convolution network going deeper.But for infrared target detection tasks,it must rely on the network to have a high-level semantic understanding of the entire infrared image.In response to these problems,this article is based on deep learning algorithms,and mainly does the following work:(1)An adaptive feature fusion network is proposed: Aiming at the specific scene of infrared small target detection,an adaptive feature fusion method is proposed.The traditional feature fusion mode mainly includes horizontal connection and top-down or down-top vertical connection.Different from the traditional structure,the adaptive fusion feature structure designed in this paper can enable the network to learn how to filter other levels of features,thereby increasing the weight of important information,suppressing secondary features,and training the optimal fusion method.Experiments have proved that the performance of the proposed new network has been significantly improved,and the precision rate,recall rate,and m AP have been improved.(2)Combining channel attention and position information,and attention structure that can use channel attention and position information at the same time is proposed and used for infrared small target detection tasks.Traditional convolutional networks often do not model the channel correlation.In recent years,researchers have proposed the attention structure for computer vision from this direction.The proposed structure can significantly improve the feature extraction of channel information by the relevant model.The method proposed in this paper combines the results of predecessors,starting from the location information that is less considered by general vision,and fusing the location information and the channel attention structure,so that the network can model the channel correlation while sensing the remote dependencies,the network performance was significantly improved by this structure.According to the experimental results,the effectiveness of the structure we proposed has been proved.The missed and false detection rate of small targets in new networks was effectively reduced without increasing too many parameters.The related experimental results have been improved.The experimental results show that the new network in this paper makes a certain improvement in the small target detection in infrared scenes.